A Highly Energy-Efficient Hyperdimensional Computing Processor for Biosignal Classification.

Area of Science:

• Biomedical Engineering
• Computer Science
• Neuromorphic Computing

Background:

• Hyperdimensional computing (HDC) is a brain-inspired paradigm for high-accuracy classification in biomedical applications.
• Existing HDC processors face energy efficiency challenges due to large memory requirements for hypervectors, scaling with sensor count.

Purpose of the Study:

• To enhance the energy efficiency of HDC processors for biosignal classification.
• To explore an on-the-fly hypervector generation method using cellular automata to reduce memory footprint.

Main Methods:

• A novel HDC architecture replacing memory with a cellular automaton for on-the-fly hypervector generation.
• Integration of vector folding for real-time classification latency.
• Post-layout simulation on an emotion recognition dataset with 200 channels.

Main Results:

• The proposed architecture achieved 39.1 nJ/prediction, a 4.9x energy efficiency improvement over state-of-the-art HDC processors.
• A 9.5x per-channel energy efficiency improvement was observed.
• At maximum throughput, a 10.7x improvement (33.5x per channel) was achieved.
• HDC demonstrated 9.5x greater energy efficiency compared to an optimized Support Vector Machine (SVM) processor for the same task.

Conclusions:

• The cellular automaton-based HDC architecture significantly improves energy efficiency for biosignal classification.
• This approach is highly effective for real-time, on-board processing, outperforming traditional SVMs.
• HDC is positioned as a leading paradigm for energy-efficient, high-accuracy biosignal classification.